The authors, Jiapeng Li and colleagues, describe the collection as the first large-scale benchmark designed specifically for fine-grained street-tree classification. Their experiments with pretrained vision models establish baselines, but also show that current systems remain challenged by the visual and statistical complexity of urban tree imagery. StreetTree addresses a data gap that has restricted progress in computer-assisted street-tree identification.

The authors identify this capability as important for urban planning, streetscape management and the assessment of urban ecosystem services. Yet the field has lacked a benchmark that is simultaneously large, geographically diverse and publicly available. The dataset includes more than 8,300 common street-tree species, creating a fine-grained classification problem with thousands of categories.

The authors supplement the image collection with expert-verified observational data. The supplied study record does not provide a single overall accuracy value for the entire dataset, nor does it identify one model as universally successful. Instead, the contribution is framed as a benchmark and baseline study: a shared body of data, a taxonomic organizing structure and experiments that reveal where current pretrained vision models remain limited.

Fine-grained tree recognition is difficult because different species can look similar in street images. The study specifically identifies high inter-species visual similarity as a challenge for pretrained models. StreetTree also contains a long-tailed natural distribution, meaning that some species are represented by many more examples than others.

Variation within a single species creates an additional test. The authors report substantial intra-class variation linked to seasonal changes, and the benchmark was assembled in real urban environments where illumination, occlusion from buildings and differing camera angles produce varied imaging conditions. A tree photographed at different times or from different positions may therefore present a markedly different image from another example of the same species.

To organize this difficult label space, StreetTree provides a hierarchical taxonomy that extends from order to family, genus and species. The hierarchy supports two related research directions: fine-grained classification and representation learning. Researchers can use the taxonomic structure to examine predictions at multiple levels, rather than treating every species as an unrelated class.

The authors also intend the taxonomy to support hierarchical classification research, in which relationships among categories can be represented explicitly and assessed as part of the learning problem. Experiments with a range of pretrained vision models establish the benchmark’s baselines and expose the limitations of existing methods under the real-world conditions represented in the data. According to the authors, the models continue to struggle with the combined effects of inter-species similarity, long-tailed class distributions, seasonal intra-class variation and diverse image capture.

The reported finding is therefore not that street-tree identification has been solved. Rather, the experiments show that current pretrained approaches do not automatically overcome the complexities of street-level imagery when the task requires species-level discrimination. StreetTree is presented as a resource at the intersection of computer vision and urban science.

Its scale, species coverage and international scope allow researchers to work with the same data when comparing classification methods, studying hierarchical representations and investigating performance under varied urban conditions. The public collection may support work related to tree mapping, planning and greener city management, as stated by the authors, but the supplied evidence does not claim that any one released model has already delivered those applications at city scale. It instead provides a foundation on which more capable and reliable methods can be evaluated.

The manuscript, titled “ArborVerse: Scaling Fine-Grained Urban Tree Species Classification to Millions of Trees Worldwide,” was posted on 21 October 2026 and is currently under review in Earth System Science Data. The StreetTree dataset is publicly available through Figshare. Because this is a preprint under discussion rather than a final journal article, its results and wording remain subject to the review process.

Even so, the project makes three core contributions immediately clear: a public benchmark containing more than 12 million images, coverage of over 8,300 street-tree species across 133 countries and five continents, and experimental evidence that today’s pretrained vision models still face substantial difficulty in recognizing street trees from unconstrained urban photographs.

Subject of Research: A global image benchmark for fine-grained street-tree species classification with pretrained vision models

Article Title: ArborVerse: Scaling Fine-Grained Urban Tree Species Classification to Millions of Trees Worldwide

Article References: Li, J., Huang, Y., Zhang, F., Tao, S., Dong, Z., Yang, J., & Liu, Y. (2026). ArborVerse: Scaling Fine-Grained Urban Tree Species Classification to Millions of Trees Worldwide. https://doi.org/10.5194/essd-2026-380

Image Credits: AI Generated

DOI: 10.5194/essd-2026-380

Keywords: StreetTree dataset, street trees, tree species, computer vision, urban ecology, image classification, machine learning, taxonomic hierarchy, urban planning, global cities, pretrained models, ArborVerse

The original research article, appearing in Frontiers in Signal Processing, presents the framework as an automated way to analyze pavement images for condition assessment and maintenance planning. The authors identify manual visual inspection as time-consuming, costly and subjective, while also noting that it can be difficult to apply across large road networks. Their system was designed to go beyond a simple crack-versus-no-crack distinction by assigning pavement imagery to six specific categories. This scope matters when interpreting the results: the reported performance concerns how well the models classified the selected image dataset, not how accurately they would assess every road under different environmental, material or camera conditions.

Methodologically, the study used transfer learning with EfficientNet-B0 and ResNet50. To address variation in pavement appearance and imbalance among the six classes, the training process included data augmentation and weighted cross-entropy loss. The researchers organized the data into training, validation and testing sets and evaluated the models using accuracy, precision, recall, F1-score and confusion-matrix analysis. EfficientNet-B0 performed best overall, with an accuracy of 89.8%, compared with 73.3% for ResNet50. Its class-level F1-scores were 0.92 for alligator cracks, 0.92 for longitudinal cracks, 0.93 for transverse cracks and 1.00 for non-cracked pavement surfaces. The supplied study abstract does not report corresponding F1-scores for branched cracks, random cracks or every other performance metric, so the full class-by-class pattern cannot be characterized from the available evidence.

The researchers also derived a Road Condition Index, or RCI, from the distribution of predicted crack categories. In their application, the index differentiated levels of pavement degradation, ranging from routine monitoring to urgent rehabilitation, and was used to support maintenance prioritization across multiple condition categories. The RCI therefore converted model classifications into an ordered assessment intended to assist decision-making. Because the index is derived from predicted categories rather than reported as a direct measurement of pavement strength, service life or repair performance, the study supports its use as a classification-based decision-support output. The reported results show that EfficientNet-B0 outperformed the tested ResNet50 configuration on the selected pavement-image data and that the proposed RCI distinguished condition categories. They do not establish that the framework is already deployed, that its classifications are universally reliable, or that prioritizing roads in this way produces measurable safety or infrastructure outcomes.

Subject of Research: Deep-learning classification of pavement cracks and model-based road-condition assessment using SMAC pavement images.

Article Title: Deep learning–based pavement crack classification and road condition assessment for intelligent transportation infrastructure

Article References: Chouiekh, C., Hammouch, W., Yahyaouy, A., & Sabri, M. A. (2026). Deep learning–based pavement crack classification and road condition assessment for intelligent transportation infrastructure. Frontiers in Signal Processing, 6, 1901600. https://doi.org/10.3389/frsip.2026.1901600

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1901600

Keywords: pavement crack classification, EfficientNet-B0, ResNet50, transfer learning, road condition index, pavement monitoring, image classification, maintenance prioritization, SMAC, Deep, learning, based

Conventional deep-learning approaches face challenges in histopathology because normal and cancerous tissue can share similar features, while cancer morphology can vary substantially within a class. A single model must also distinguish normal tissue from several cancer categories at once, producing a prediction without an explicit breakdown of the decisions involved. The proposed framework first separated normal tissue from cancerous tissue, then directed each image to a specialized classifier for normal-tissue categorization or cancer-subtype identification.

The researchers independently evaluated candidate deep-learning architectures at each stage and selected the model suited to each particular classification task. This design allowed different stages to focus on narrower decisions. The framework was developed and evaluated on LC25000 before being assessed on LungHist700, an independent dataset with different image characteristics and class distributions. The difference between the two accuracy results shows that performance measured on one benchmark did not fully transfer to the separate evaluation setting.

The study also used four interpretability techniques—Grad-CAM, Grad-CAM++, LIME and occlusion sensitivity—to identify image regions contributing to the hierarchical predictions. These tools provided complementary views of the model’s decision process while the framework retained its staged classification structure. The reported results support the combination of strong benchmark performance, task-specific model selection and complementary explanation methods, but the analysis concerns computational image classification rather than patient diagnosis.

The findings remain bounded by the available data and evaluated setting. The study used two histopathology datasets and addressed five tissue categories, and its abstract reports overall accuracy without establishing equivalent performance across broader clinical conditions or patient populations. It also does not show that the highlighted image regions represent medically valid biological features or that the framework would improve diagnostic outcomes. Accordingly, the results demonstrate the potential of hierarchical decision decomposition and explainable deep learning for computer-aided histopathology classification, while further validation would be needed to establish performance in clinical use.

Subject of Research: Explainable hierarchical deep learning for computational lung and colon histopathology classification

Article Title: An interpretable hierarchical deep learning framework for lung and colon cancer detection

Article References: Talukder, M. A., Rafeh, R., Chu, N. H., & Uddin, M. A. (2026). An interpretable hierarchical deep learning framework for lung and colon cancer detection. Scientific Reports. https://doi.org/10.1038/s41598-026-75965-4

Image Credits: AI Generated

DOI: 10.1038/s41598-026-75965-4

Keywords: deep learning, histopathology, lung cancer, colon cancer, explainable AI, medical imaging, computer vision, hierarchical classification, artificial intelligence, image classification, interpretable, hierarchical

The study used a publicly available Kaggle dataset containing images divided among four classes: Black Spot, Canker, Fresh and Greening. It evaluated eight fixed feature extractors—seven convolutional neural networks and a ViT-S/16 Vision Transformer—together with four classical classifiers: SVM, k-nearest neighbors, random forest and XGBoost. All image models were evaluated using the same dataset splits, preprocessing and metrics under 5 × 5 repeated stratified cross-validation, comprising five repeats of five folds with an 80:20 class-stratified train-test ratio in each fold.

To prevent information from held-out images entering the dimensionality-reduction stage, researchers fitted PCA separately on the training-fold features in every iteration and then applied the resulting transformation to the corresponding test fold without refitting it. The researchers used a target of retaining 98% of feature variance. A sensitivity analysis found that mean accuracy plateaued across the 95%–98% range and did not improve at 99%, supporting the selected threshold.

Across the 32 backbone-classifier combinations, SVM was consistently the strongest or nearly strongest classifier after PCA reduction. For ViT features, SVM significantly outperformed k-nearest neighbors, random forest and XGBoost, with all reported comparisons having p < 0.001. Paired tests across the 25 matched partitions also found that the ViT + PCA + SVM combination significantly outperformed DenseNet201 + PCA + SVM (p < 0.001) and EfficientNetB0 + PCA + SVM (p = 0.004).

The class-level results for the proposed ViT pipeline were not limited to the two easiest categories. Aggregated over the 25 held-out folds, F1-scores were 97.70% for Black Spot and 97.63% for Canker, compared with 99.97% for Fresh and 100.00% for Greening. The two visually similar disease classes therefore had somewhat lower, though still high, F1-scores than the other categories.

Ablation tests indicated that PCA’s main measured contribution was feature reduction rather than an increase in accuracy. For ViT, retaining 98% of variance required an average of 170 of 384 components, a reduction of about 55.7%. ViT + PCA + SVM was statistically indistinguishable from ViT features without PCA (99.01% ± 0.62%; p = 0.16) and from SVM applied to the full 384-dimensional feature space (99.12% ± 0.69%; p ≈ 1.0). The corresponding ablation for DenseNet201 produced a larger reduction, about 76.6%, while the complete pipeline did not significantly differ from unreduced DenseNet201 features (p = 0.25).

Explainability analyses used occlusion sensitivity and SHAP for the ViT configuration, while DenseNet201 was assessed with Grad-CAM and SHAP because the study states that Grad-CAM is not directly applicable to transformer attention. ViT occlusion maps showed high sensitivity in localized fruit-surface regions for Fresh and Canker, distributed sensitivity across the lower half of the fruit for Greening, and sensitivity around visible dark spots for Black Spot. SHAP summaries showed contributions distributed across multiple principal components, with both positive and negative values. These representative examples indicated reliance on fruit-surface evidence, although some ViT maps also showed sensitivity near image borders.

The computational benchmark also distinguished accuracy from model size. ViT had 21.67 million parameters and an approximate size of 82.6 MB, while DenseNet201 had 18.32 million parameters and an approximate size of 69.9 MB. The measured mean inference time was 99.07 ms for ViT and 118.66 ms for DenseNet201. EfficientNetB0 and MobileNetV2 were substantially smaller, with 4.05 million and 2.26 million parameters, respectively.

The findings are bounded by the dataset’s modest size and single public source. The 1,090 field or orchard photographs were whole intact fruit, but their repository did not provide original per-image acquisition details such as camera device, precise orchard location or controlled-versus-field capture conditions. The images also may not represent the full variation in cultivars, growing regions, lighting and sensors encountered elsewhere. The authors therefore characterized the reported accuracies as indicative rather than conclusive and identified validation on independently collected, multisource imagery as necessary future work.

Subject of Research: Experimental benchmarking of PCA and SVM classifiers with deep image features for four-class orange fruit disease classification

Article Title: A PCA-based deep feature optimization framework for explainable orange fruit disease classification

Article References: Hingmire, A., Golande, A., Bhutnal, V., Pande, S. D., & Pande, T. (2026). A PCA-based deep feature optimization framework for explainable orange fruit disease classification. BMC Plant Biology, 26(1), Article 1746. https://doi.org/10.1186/s12870-026-09984-8

Image Credits: AI Generated

DOI: 10.1186/s12870-026-09984-8

Keywords: orange diseases, image classification, Vision Transformer, principal component analysis, support vector machine, explainable AI, cross-validation, DeepNet201, computer vision, fruit disease dataset, PCA-based, deep

The proposed architecture used an ensemble image encoder comprising ResNet50, DenseNet121 and EfficientNet-B0, together with dedicated clinical and genomic encoders. An adaptive fusion mechanism combined modality-specific predictions through convex weights driven by predictive entropy estimated with Monte Carlo Dropout. In tests where auxiliary modalities were uninformative, the framework was reported to suppress their influence and maintain identical test-set predictions across all image-containing configurations. A static equal-weight baseline, by comparison, had an AUC of 0.982 rather than 0.993.

The framework also produced well-calibrated probabilities in the reported evaluation, with a Brier score of 0.025 and expected calibration error of 0.014. Confidence-based selective prediction was described as showing potential for human-in-the-loop triage. This is a measured evaluation result, but the abstract does not establish improved clinical outcomes or demonstrate benefit in prospective screening.

Important limitations constrain the scope of the work. The study was explicitly presented as a proof-of-concept, and paired patient-level multimodal data were unavailable. Clinical features were therefore derived from an independent risk-factor cohort, while genomic features were synthetically generated. Cross-dataset transfer experiments showed a generalization gap, indicating that further external validation is needed. The reported performance came from cervical-cell image benchmarks and should not be interpreted as evidence that the full multimodal framework has been validated for routine patient screening.

Subject of Research: Proof-of-concept evaluation of uncertainty-aware multimodal classification of cervical-cell images

Article Title: Uncertainty-aware adaptive fusion with graceful degradation for trustworthy cervical cytology

Article References: Chaabane, S. B., Bushnag, A., Oyouni, A. A. A., Massoudi, W., Abuzneid, S., & Aluneizi, S. (2026). Uncertainty-aware adaptive fusion with graceful degradation for trustworthy cervical cytology. Scientific Reports. https://doi.org/10.1038/s41598-026-73797-w

Image Credits: AI Generated

DOI: 10.1038/s41598-026-73797-w

Keywords: cervical cytology, deep learning, multimodal fusion, predictive uncertainty, image classification, model calibration, synthetic genomic data, external validation, Uncertainty-aware, adaptive, fusion, graceful

The problem is known as spurious correlation: a feature that is not genuinely predictive of the target becomes associated with it in the training data. A bird classifier might learn that water backgrounds indicate one species and land backgrounds indicate another, while a face classifier might associate blond hair with one gender. Such patterns can produce high average accuracy during training and testing while masking serious failures in underrepresented combinations. When a new example breaks the accidental relationship, performance can fall sharply. This makes average accuracy alone an incomplete measure of whether a model has learned the intended task.

The study builds on UnLearning from Experience, or ULE, a method introduced by the same research group. ULE trains two neural networks in parallel. The first, called the biased model, learns through standard empirical risk minimization, meaning it minimizes average training loss in the conventional way. The second, called the debiased model, is trained to solve the same classification task while using different features. It is encouraged to do so through a loss based on the gradients, or saliency maps, of the two networks: the debiased model receives a signal to direct its attention away from the features emphasized by the biased model.

That design is motivated by the “simplicity bias” hypothesis, which holds that neural networks tend to learn easier features before more complicated ones. If shortcut features are easier to learn than the task-relevant characteristics of birds, cars, faces or language, an early-stage network may be especially likely to use them. The researchers calculate the gradient of each model’s output with respect to its input and use the mean squared error between the negated biased-model gradient and the debiased-model gradient as a feature-saliency term. A separate cross-entropy term still requires the debiased network to predict the correct label, while a hyperparameter called lambda controls the balance between classification and saliency matching.

Weighted UnLearning from Experience, abbreviated wULE, changes how this combined loss is applied to individual examples. Earlier ULE treats every training item alike, implicitly assuming that all contain a shortcut. That can be counterproductive: forcing the two models to diverge on examples with no spurious correlation may prevent the debiased model from learning useful behavior that the biased model has already captured correctly. The new formulation therefore introduces two positive scaling factors. Alpha can increase the weight assigned to the classification loss for estimated non-spurious examples, while beta can increase the saliency loss for examples likely to contain spurious features. The authors describe wULE as a generalization of ULE, for which the two new factors can reduce to the original weighting scheme.

Because ground-truth group annotations are generally unavailable in practical use, the researchers tested two ways of estimating which training examples contain spurious correlations. A JTT-based method treats examples correctly classified by the biased network after an early training epoch as the spurious set, on the assumption that shortcut examples will already have been learned. A SPARE-based method instead clusters the biased model’s logits and assumes that the majority cluster represents examples with spurious correlations. In both cases, the biased network is frozen after the selected epoch and the debiased model continues training. Experiments on Waterbirds showed that both estimators worked best when the biased model was frozen early. The SPARE-based estimate was consistently stronger than the JTT estimate, producing better worst-group accuracy and remaining more useful at later freezing points, after the biased model had begun fitting the training set more comprehensively.

The evaluation covered Waterbirds, CelebA, UrbanCars and MultiNLI. Waterbirds tests whether a classifier can identify landbirds and waterbirds when backgrounds do not match the usual correlation. CelebA asks models to predict blond or non-blond hair when gender is correlated with the label. UrbanCars classifies cars as urban or country while varying both background and a co-occurring object, allowing researchers to test simultaneous shortcuts. MultiNLI is a natural-language inference task involving entailment, neutral and contradiction relationships; negation words in hypotheses were treated as a potentially spurious attribute. The principal metric was worst-group accuracy, which records performance in the least successful subgroup, alongside overall average accuracy. Group labels were used for evaluation and hyperparameter selection, not to identify spurious examples during wULE training.

On Waterbirds, both estimated variants of wULE significantly improved worst-group accuracy over ULE and approached results obtained by an “oracle” version that used ground-truth information to identify spurious examples. Across the broader comparison with previously published methods, wULE ranked among the strongest approaches: the authors placed it second on Waterbirds, UrbanCars and MultiNLI, and fourth on CelebA when using the same ResNet-50 backbone used for like-for-like comparisons. The SPARE-based variant produced the reported wULE results in that comparison. Its worst-group and average accuracies were also relatively close, which the authors interpret as evidence that it was less dependent on majority-group shortcuts than methods showing a larger gap between those measures. These are benchmark results rather than evidence of performance in uncontrolled production settings.

Additional analyses examined how the method responds to errors, model capacity and the choice of loss. Randomly corrupting the estimated spurious set caused worst-group accuracy to decline, but at corruption levels resembling the estimation errors observed in practice, performance remained near the results obtained by the two practical estimators. Increasing the emphasis on non-spurious examples through alpha and on saliency separation for spurious examples through beta generally improved results in the tested range, while the saliency term received somewhat more weight than classification. Replacing mean squared error with mean absolute error slightly reduced worst-group accuracy but marginally increased average accuracy, so the researchers retained mean squared error for their main analysis. Tests combining ResNet-18, ResNet-50 and ViT-H-14 backbones showed that a more expressive debiased model generally improved performance, whereas the biased model’s architectural complexity had less influence.

Explainable-AI heatmaps provided a qualitative view of the learned behavior. On UrbanCars, the standard biased model often assigned high saliency to the background or the accompanying object as well as to the car. Models trained with wULE concentrated more strongly on the car and gave less attention to the two misleading attributes. On Waterbirds, the debiased model increasingly focused on the bird as training proceeded, while the biased model continued to emphasize the background. The researchers also showed that the method need not carry an extra inference-time burden: both networks are trained in parallel only for the first k epochs, the biased model is then discarded, and only the debiased model is used for predictions. Overall, the results support weighted treatment of likely shortcut examples as a practical extension of ULE, while the reliance on early-training behavior, estimated groups, pretrained backbones and benchmark-specific tuning remain important limits on how broadly the findings can be generalized.

Subject of Research: Machine-learning method for improving robustness to spurious correlations in image and language models

Article Title: Weighted UnLearning from Experience for Improved Spurious Correlation Robustness

Article References: Mitchell, J., Martinez-del-Rincon, J., & McLaughlin, N. (2026). Weighted UnLearning from Experience for Improved Spurious Correlation Robustness. Machine Learning, 115(11), Article 251. https://doi.org/10.1007/s10994-026-07164-1

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07164-1

Keywords: spurious correlations, machine learning, robust AI, deep learning, out-of-distribution generalization, UnLearning from Experience, image classification, natural language inference, gradient saliency, benchmark datasets, Weighted, UnLearning

The researcher behind the study, Xiaofan He of the College of Art and Design at Jiangsu University of Technology, identified a persistent gap in existing cultural virtual display systems: computer vision pipelines that classify heritage images rarely consider how their outputs could guide navigation, while path planning methods in human-computer interaction seldom exploit high-precision classification to personalize tours. Existing systems either achieve accurate classification but offer interaction paths unrelated to visitor interests, or provide smooth experiences built on coarse content labels. The new work closes that loop by feeding classification results directly into the cost function of a path planner.

The classification engine is built on a Graph Capsule Neural Network, or CapsGNN, a hybrid of capsule networks and graph neural networks. Capsule networks, unlike conventional convolutional neural networks, represent features as vectors whose length encodes the probability that a feature exists and whose orientation encodes its pose, such as rotation, scale, and deformation. This matters for Yi textiles, which are full of rotationally symmetric, repeating units and nonlinear deformations. Standard CNN pooling discards precise spatial position information, so the same pattern viewed at different angles is easily misclassified. Capsules, routed through a dynamic routing algorithm, can recognize what the study calls different poses of the same type.

The key architectural modification replaces the pose matrix in the original CapsGNN with a Graph Convolutional Network (GCN). The pose matrix parameterization in standard CapsGNN struggles to exploit adjacency information when processing graph structures, making it hard to capture the spatial relationships between repeating units in fabric patterns. GCNs, by contrast, explicitly aggregate neighbor node features through Laplacian smoothing operations, integrating low-level capsules into higher-level representations via neighborhood-weighted aggregation. An attention mechanism further refines the weighting of capsule information. The resulting CapsGNN-GCN model passes Yi fabric images through input, initial capsule, hierarchical capsule, and class capsule layers, where low-level capsules cluster into high-level capsules that map to specific pattern categories.

To train and test the model, the study assembled a dataset of 3,086 Yi fabric images collected online, from books, and through field photography, spanning five labeled categories: 700 abstract geometric images, 570 fire patterns, 460 horn patterns, 610 bracken patterns, and 746 other traditional patterns. Data augmentation through brightness adjustment, rotation, horizontal flipping, random cropping, and scaling expanded each category to 4,500 images, all resized to 300 by 300 pixels. Under five-fold cross-validation, CapsGNN-GCN reached a training accuracy of 96.8 percent and an F1 score of 0.93. Ablation tests confirmed the contribution of each module: removing the GNN module dropped training accuracy to 83.4 percent, removing GCN lowered it to 87.6 percent, and removing both collapsed test accuracy to 70.5 percent.

Head-to-head comparisons against ResNet50 V2, EfficientNet B0, and a GCN-CNN baseline showed the improved model winning every category. In the Abstract Geometry class, CapsGNN-GCN achieved 96.6 percent accuracy with only 10 misclassifications, while the Sheep’s Horn Pattern and Other Traditional Patterns categories reached 95.7 and 95.5 percent respectively. ResNet50 V2 managed only 87.6 percent on fire patterns, and EfficientNet B0 topped out at 95.5 percent on abstract geometry. Paired t-tests confirmed the differences were statistically significant at p less than 0.01. Notably, when evaluated on the original unaugmented test set, the model’s accuracy fell by only 0.6 percentage points, to 91.7 percent, indicating the augmentation had not inflated performance. The few failures clustered in texture-similar regions between fire and ram’s horn patterns and between fern and abstract geometry designs, accounting for just 0.7 percent of the test set.

The second half of the system tackles how visitors move through virtual heritage spaces. Using oblique photography and 3D modeling software integrated into Unreal Engine 4, the study reconstructed a village in Liangshan as a navigable virtual scene, gridded into a 100 by 60 cell raster over an 800 by 1,000 meter area. The team then modified the classic A* pathfinding algorithm, which combines best-first search heuristics with Dijkstra-style guarantees of shortest paths, by adding three user experience constraints to its heuristic function: landmark salience, user movement trajectory overlap, and user preference. Landmark salience combines visual contrast with semantic importance ratings from three cultural experts, diffused across the grid with a Gaussian kernel. Trajectory overlap was computed from the free-exploration walking paths of 30 users, and preference grids were linked to the classification model’s pattern labels, scored on a five-point Likert scale. A cubic uniform B-spline curve then smooths the raw grid path to reduce sharp turns.

Across six start-end pairs connecting key village nodes such as the center, entrance, sacrificial site, handicraft workshop, viewing platform, and cultural square, the improved algorithm produced an average total path length of 523.7 plus or minus 42.9 meters, roughly 6.3 to 9.6 percent shorter than simulated annealing and dynamic window approach baselines. The average number of turns fell to 13.7 plus or minus 2.2, the average turning angle dropped to 6.5 plus or minus 0.6 degrees, a 30 to 40 percent reduction in corner severity, and the average computation time was 15.3 plus or minus 1.9 seconds, well below the 19.5 seconds of traditional A* and the 25.4 seconds of DWA. Sensitivity analysis showed satisfaction scores changed by no more than 0.3 points on a five-point scale when weights were perturbed, indicating the algorithm is robust to weight selection.

Human trials with 50 participants, randomized between the new system and a non-interactive static display platform, demonstrated substantial gains in experience quality. Task completion averaged five minutes with a system response time of 0.5 seconds, immersion scored 4.2 on the Slater Immersion Scale, satisfaction reached 4.5 on a five-point Likert scale, and the operation error rate was just 3 percent. Independent samples t-tests showed the immersion and satisfaction differences were significant at p less than 0.001 with large effect sizes. A second experiment with 30 participants from diverse cultural backgrounds found that knowledge of Yi culture increased by 25 percent after using the system, with pre-test scores rising from 24 to 30 points, compared with a 10 percent gain for the traditional platform, and the cultural identity score reached 4.3.

The study’s broader significance lies in its methodological transferability. The quantification of landmark salience relies only on visual contrast and expert semantic scoring, the trajectory overlap grid can be generated from exploration data in any virtual environment, and the preference grid links directly to any classification model with a label-to-preference mapping, meaning the framework could extend to museums, educational platforms, and other heritage sites with spatial navigation needs. The author acknowledges limitations: the experiments covered only Yi fabric images and one virtual village scene, and adaptability to other cultural elements remains unverified. Future work aims to expand the method to other heritage types and optimize real-time feedback, but the demonstration that human-computer interaction factors can be quantitatively embedded in a path planning cost function marks a concrete step toward virtual heritage systems that do not just display culture, but anticipate what visitors want to see.

Subject of Research: A virtual display experience system for Yi cultural heritage that integrates CapsGNN-based fabric image classification with user-experience-driven A* path planning in a virtual village

Article Title: Virtual display experience design for Yi cultural heritage using CapsGNN and user experience

Article References: He, X. (2026). Virtual display experience design for Yi cultural heritage using CapsGNN and user experience. Discover Artificial Intelligence, 6(1), Article 1193. https://doi.org/10.1007/s44163-026-02017-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02017-1

Keywords: CapsGNN, graph capsule neural network, Yi culture, cultural heritage, virtual reality, image classification, path planning, A* algorithm, user experience, Unreal Engine, digital preservation, Liangshan

The research was conducted during the 2023 Kharif season at an experimental field operated by Praanadhaara Organised Agro Forestry Private Limited in Bapatla District, Andhra Pradesh, a tropical region where monsoon rainfall makes rice the dominant crop. The team, drawn from institutions including Acharya N. G. Ranga Agricultural University, the ICAR-Indian Institute of Rice Research, and South Dakota State University, cultivated three plots of roughly two acres each: broadcasted direct seeded rice, wet direct seeded rice, and transplanted rice. Each method leaves a distinct signature in the field, from the scatter pattern of broadcast seed to the flooded, orderly rows of transplanted seedlings, but capturing that signature requires imagery far sharper than what orbiting satellites can provide.

Using a DJI Mavic 2 Pro drone equipped with a 20-megapixel Hasselblad camera, the researchers flew automated waypoint missions on three dates: October 15, 2023, October 27, 2023, and December 1, 2023. Flights were held to a constant altitude of 40 meters, with 80 percent forward and 75 percent side image overlap, under clear skies and gentle winds between 3 and 4 meters per second. Each mission produced roughly 425 raw red-green-blue images at a ground sampling distance of 1.1 centimeters per pixel. That resolution is extraordinary by remote sensing standards: previous efforts to map direct seeded rice relied on satellite sensors with resolutions between 10 and 250 meters, which blur out the fine structural cues that separate one planting method from another.

After the raw images were orthorectified into georeferenced mosaics, the team overlaid GPS-derived plot boundaries using QGIS software and cut the mosaics into 400-by-400-pixel patches. The resulting dataset contained 1,966 labelled patches: 672 of broadcasted direct seeded rice, 612 of transplanted rice, and 682 of wet direct seeded rice. Because all three systems were managed under identical agronomic conditions and sowing schedules were controlled, any differences the models detected could be attributed to the establishment method itself rather than to confounding variation in crop age or management.

The classification pipeline paired a convolutional neural network with a conventional machine learning classifier. ResNet-50, a 50-layer deep residual architecture, served as the feature extractor, converting each image patch into a 2,048-dimensional vector that encodes spatial and textural information. The authors chose ResNet-50 deliberately: its transfer learning capability and moderate complexity suit relatively small drone datasets, where deeper networks would risk overfitting. Those feature vectors were then fed to six different classifiers: k-nearest neighbor, support vector machine, decision tree, random forest, Naïve Bayes, and a neural network, producing six hybrid models. Two additional end-to-end architectures, YOLOv5 and YOLOv8, were trained on the same data for comparison.

Performance was tested rigorously. The dataset was split across nine different training-to-testing ratios, from 10:90 up to 90:10, and evaluated with overall accuracy, precision, recall, F1-score, the kappa coefficient, and mean average precision. The clear winner was the ResNet-50 plus neural network combination, which achieved average overall accuracies of 93.16 percent, 95.81 percent, and 93.31 percent across the three acquisition dates. No other model matched that consistency. The next best, ResNet-50 plus support vector machine, reached 92.07, 95.24, and 92.09 percent, while the weakest performers, Naïve Bayes and decision tree hybrids, hovered in the high 70s. The YOLO models, despite their popularity in agricultural computer vision, landed in the middle at roughly 78 to 89 percent.

The winning model also posted the strongest secondary metrics, with precision between 0.93 and 0.97, recall between 0.92 and 0.97, F1-scores from 0.92 to 0.97, and kappa coefficients of 0.89 to 0.96. The kappa coefficient is particularly telling, because it corrects for agreement that could occur by chance, meaning the model was genuinely discriminating between classes rather than exploiting class imbalance. Notably, the peak performance came at the second acquisition date, when direct seeded rice was mid-establishment, suggesting the features that distinguish planting methods become most legible partway through the season.

The authors are candid about the study’s limitations. Because transplanted rice was not put into the ground until November 3, the first two drone flights captured that treatment only as pre-transplant field preparation, so early-stage classifications partly reflect soil, water, and land preparation differences rather than a mature transplanted canopy. The design also synchronized sowing dates across treatments, whereas real-world farms stagger planting according to rainfall, water availability, and labor. At regional scales, such phenological variation could complicate classification, and the authors recommend validating the framework across multiple seasons, planting dates, and regions before operational deployment.

Even so, the implications are substantial. Knowing which establishment method dominates a given landscape matters for precision agriculture: transplanted rice, wet direct seeded rice, and broadcast rice differ sharply in water consumption, labor demands, methane emissions, and input requirements. Accurate, fine-grained maps of establishment methods could sharpen area estimates, guide site-specific management, and support policies promoting water-saving direct seeding. The authors point toward future upgrades, including multispectral and hyperspectral sensors to capture plant physiology, transformer and self-supervised architectures that need less labelled data, explainable AI for transparency, and lightweight models that run in real time aboard the drone itself. For now, the study stands as a compelling proof of concept: an off-the-shelf drone, a pretrained network, and a well-chosen classifier can read the hidden history of how a rice field was planted, written in centimeter-scale patterns of leaves, soil, and water.

Subject of Research: Deep learning classification of rice establishment methods using high-resolution drone imagery

Article Title: Deep learning-based classification of wet direct seeded rice, broadcasted direct seeded rice and transplanted rice using drone imagery for precision agriculture

Article References: Gubbala, A. L., Rathod, S., Dasyam, R., Rapolu, M. K., Dasari, A. K., Kurra, P., Madireddy, H. R., Shashishekhar, P. R., Mural, R. V., Kumar, A., & Sundaram, R. M. (2026). Deep learning-based classification of wet direct seeded rice, broadcasted direct seeded rice and transplanted rice using drone imagery for precision agriculture. Frontiers in Remote Sensing, 7, Article 1698781. https://doi.org/10.3389/frsen.2026.1698781

Image Credits: AI Generated

DOI: 10.3389/frsen.2026.1698781

Keywords: deep learning, drone imagery, precision agriculture, remote sensing, ResNet-50, YOLOv8, rice, direct seeded rice, transplanted rice, UAV, image classification, crop mapping

The core dilemma the review describes is architectural. The most accurate image-analysis models are built by centralizing data: collecting every image in one place so a neural network can learn from all of it. But centralizing surveillance footage concentrates exactly the kind of sensitive biometric and behavioral information that privacy laws, governance policies, and re-identification research worry about. Federated learning, introduced as a practical alternative in 2017, inverts the arrangement. Instead of shipping images to a central server, each camera or client trains a local copy of the model on its own data and sends only model updates—the gradients that describe how its parameters should change—to an aggregator, which averages them into a shared global model using the FedAvg algorithm. Raw footage never leaves the device. The scheme fits urban sensing naturally, because cities already have enormous numbers of distributed devices with local storage and limited bandwidth.

Yet federated learning alone, the review emphasizes, does not actually guarantee privacy. This is the surprise that transformed the field. In 2019, researchers demonstrated that the gradients a client uploads can be mathematically inverted to reconstruct training images pixel-for-pixel, in an attack memorably titled Deep Leakage from Gradients. Successive refinements made reconstruction feasible for realistically sized batches, extended the technique to whole image sets, and added recursive linear-algebraic attacks. A 2023 systematic categorization of the field divides gradient-leakage methods into bias attacks, optimization-based attacks, and linear equation-solver attacks, and confirms something especially worrying for surveillance: images are substantially more vulnerable to near-exact reconstruction than text or tabular data, because they are high-dimensional and redundant. Beyond gradient inversion, the review catalogues membership inference attacks that reveal whether a particular person’s image was in a training set, property inference attacks that deduce aggregate attributes of a client’s data, and model extraction attacks that steal the trained model itself. In other words, sharing updates instead of images is a necessary step, not a sufficient one.

That is where differential privacy enters. Differential privacy is a formal, mathematically defined guarantee: it bounds how much the learning process can be influenced by any single example in the training data, ensuring that the final model would look statistically similar whether or not any one image was included. In practice, the guarantee is implemented through an algorithm called DP-SGD, first benchmarked on image classification in 2016. DP-SGD computes gradients for each training example individually, clips each one to a bounded maximum size, and then adds calibrated Gaussian noise before the results are aggregated. The strength of the protection is governed by a privacy budget, denoted by the Greek letter epsilon. A smaller epsilon means stronger privacy but more noise and therefore lower accuracy—a fundamental trade-off that no amount of clever engineering has eliminated. Secure aggregation protocols are typically layered on top, using cryptography so the server sees only the noised sum of all clients’ updates and never any individual contribution. The combination of these ingredients is DPFL.

The review’s most striking quantitative finding concerns how fast this trade-off has improved. On CIFAR-10, the standard image-classification benchmark used across nearly a decade of research, the first DP-SGD results in 2016 reported 73.0 percent top-1 accuracy at a privacy budget of epsilon equal to 8, roughly 22 percentage points below a non-private baseline of about 95 percent. By 2022, careful scaling and tuning of large over-parameterized models had pushed the private state of the art to 81.4 percent at the same budget, narrowing the gap to roughly 13 to 14 points. Progress was not perfectly monotonic—the earliest result used pre-trained convolutional layers, which changes the effective privacy setup—and the trajectory has been powered by innovations such as adaptive and automatic clipping, which remove the need for hand-tuning the clipping threshold, and architecture designs optimized specifically for private training. The review also notes the cost of tighter protection: on the same architecture, accuracy falls from 73 percent at epsilon 8 to 67 percent at epsilon 2, so cities with stricter privacy requirements should expect to pay more in model performance.

Crucially, the authors draw a line that much of the broader literature blurs. Those celebrated benchmark numbers all come from centralized training on curated datasets, where a single trusted party holds everything. Federated DPFL adds a second layer of difficulty: gradients are clipped and noised locally, then aggregated across clients whose data distributions differ substantially—so-called non-IID data—because cameras in different locations see different scenes. Studies in this federated regime report smaller but more realistic gains. Clustered federated learning with adaptive local differential privacy improved the privacy-utility balance on heterogeneous Internet-of-Things data; adaptive noise mechanisms beat fixed-noise baselines at the same privacy level; and one reliable federated scheme improved accuracy by 1.1 percent while cutting convergence time by more than 9 percent. Evidence from medical imaging, which shares core mechanics with surveillance, suggests acceptable performance at moderate budgets around epsilon 10 but substantial degradation under strict budgets around epsilon 1.

Applied smart-city systems are now demonstrating what hybrid designs can achieve. A federated anomaly-detection approach for CCTV surveillance avoided sharing raw footage entirely and achieved 95 percent accuracy under random data splits, or 86 percent under balanced splits. A remote-sensing classification framework designed to tolerate a dishonest majority of malicious participants saw accuracy fall from 92.5 percent with 20 percent attacker participation to 81.84 percent with 80 percent—an eleven-point penalty that quantifies the price of distrust. Perhaps most eye-catching are the hybrid stacks: a framework combining differential privacy with homomorphic encryption, which allows computation on encrypted data, retained 95.2 percent accuracy even at a strict privacy budget of epsilon 1.0 with minimal degradation. A companion approach pairing homomorphic encryption with dynamic differential privacy and compression reached 92.6 percent accuracy while cutting transmission volume by 40 percent and latency by 43 percent. And a quantum-inspired aggregation scheme reported 94.7 percent accuracy with 43 percent less overhead and only a 2.1 percent degradation relative to a non-private baseline. A blockchain-backed framework handling 100 clients with 15 percent malicious participants reported more than 90 percent accuracy, 32 percent lower overhead, and less than 4 percent accuracy loss at epsilon 1.

The authors insert an important dose of skepticism here: those applied figures are not directly comparable to one another or to the CIFAR-10 benchmarks. They use different datasets, architectures, privacy accounting methods, and threat models, and some easier task framings—such as binary anomaly detection rather than full multi-class classification—likely inflate apparent accuracy. Nor are epsilon values necessarily on a common scale across studies, since neighboring datasets can be defined differently, as adding or removing one image versus one client, and cumulative privacy loss across training rounds further erodes the guarantee over a full deployment. A single reported epsilon can therefore carry very different real-world implications depending on context. The review concludes that no standardized guidance exists for what privacy budget meaningfully protects citizens while preserving public-safety-relevant accuracy in surveillance settings.

That gap leads to the review’s central caution: DPFL is a credible, rapidly maturing framework rather than a finished solution. Nearly all gradient-leakage attacks have been demonstrated in controlled settings with known architectures and small batches, so their success rate against production-scale systems defended by secure aggregation remains an open empirical question. Real camera networks bring non-IID data, heterogeneous hardware, and communication constraints that curated benchmarks simply do not capture. Hybrid stacks that deliver the best accuracy also add system complexity and latency that may clash with real-time public-safety requirements. The authors’ proposed path forward is fourfold: benchmark DPFL on heterogeneous, non-IID camera data rather than curated datasets; validate resistance to gradient leakage against production-scale secure aggregation and adversarial client populations; converge on domain-specific epsilon guidance balancing public-safety accuracy against re-identification risk; and evaluate hybrid privacy stacks for the latency and complexity costs they impose, not only the accuracy they recover. For a field racing to keep cities intelligent without making their citizens transparent, that measured verdict may be the review’s most valuable contribution.

Subject of Research: Differentially private federated learning for privacy-preserving image analysis in smart-city surveillance systems

Article Title: Differentially private federated learning for image analysis in smart-city surveillance

Article References: Pandya, N., & Mistry, D. (2026). Differentially private federated learning for image analysis in smart-city surveillance. Frontiers in Imaging, 5, Article 1966936. https://doi.org/10.3389/fimag.2026.1966936

Image Credits: AI Generated

DOI: 10.3389/fimag.2026.1966936

Keywords: differential privacy, federated learning, DP-SGD, gradient leakage, smart-city surveillance, image classification, privacy-utility trade-off, homomorphic encryption, secure aggregation, remote sensing, CIFAR-10, urban IoT

The study, published in the International Journal of Machine Learning and Cybernetics, was led by Anish Kumar Vishwakarma and Rakesh Ranjan, who set out to fix a structural weakness in how deep learning has been applied to agricultural image classification. Most existing pest recognition systems are built as multi-stage pipelines: one model extracts features from an image, and a separate classifier makes the final decision. This division of labor sounds tidy, but it carries real costs. Each stage adds computational overhead, and because the feature extractor is not trained jointly with the classifier, the learned representations can miss the subtle visual cues that separate one pest species from another. For insects that differ only in the pattern of wing veins or the shade of a marking on the thorax, that loss of discriminative detail is exactly where mistakes happen.

Vishwakarma and Ranjan’s answer is an end-to-end architecture built on a modified MaxViT backbone, a vision transformer design first introduced at the European Conference on Computer Vision in 2022. MaxViT is unusual among transformers because it combines two kinds of attention in a single network: window attention, which captures fine-grained local relationships within small patches of the image, and grid attention, which lets distant regions of the image exchange information. This multi-axis attention is what allows the model to notice both the texture of an insect’s body and its overall shape and position in the frame, two kinds of evidence that classical convolutional networks often handle separately.

The key innovation, however, lies in what the authors call a Multi-Stage Feature Fusion Head. Deep networks process images hierarchically: early layers detect simple edges and textures, middle layers assemble those into parts and patterns, and deep layers encode abstract semantic concepts. In many architectures, the low-level texture information is effectively discarded by the time the network reaches its final decision. The fusion head explicitly reintegrates these complementary levels, combining low-level texture features with high-level semantic features so that the final classification draws on the full spectrum of visual evidence. For visually similar pest categories, this fusion proved decisive, since fine surface texture often carries the discriminating signal that abstract features alone cannot supply.

Training such a system on a specialized agricultural dataset presents its own challenge: there is simply no jute pest corpus large enough to train a transformer from scratch. The researchers therefore employed a two-phase transfer learning strategy, starting from representations pretrained on ImageNet, the massive general-purpose image database that has become the standard starting point for computer vision. In the first phase, the pretrained backbone is adapted to the general characteristics of the target domain, stabilizing optimization before the model is asked to master the finer distinctions among pest species. In the second phase, the network fine-tunes on the jute pest data itself, allowing the ImageNet-learned features to reshape around the specific visual grammar of insects on leaves and stems. This staged approach avoids the instability that can occur when a large pretrained model is abruptly exposed to a narrow new domain.

The evaluation was carried out on the publicly available Jute Pest Dataset hosted in the UCI Machine Learning Repository, which covers seventeen jute pest species. The results were striking across every standard metric: overall classification accuracy of 98.43 percent, precision of 98.28 percent, recall of 98.28 percent, and an F1-score of 98.26 percent. The tight agreement among these numbers matters, because it indicates the model performs consistently across all seventeen classes rather than excelling on a few common pests while failing on rare ones. To guard against the possibility that a single favorable train-test split had flattered the results, the authors also ran five-fold stratified cross-validation, and the model’s performance held steady across different data partitions, confirming genuine generalization rather than memorization.

The comparison against established baselines was equally one-sided. The proposed framework consistently outperformed VGG16, ResNet50, DenseNet201, the original Vision Transformer, the Swin Transformer, and ConvNeXt, a lineup that spans a decade of computer vision architecture design. It also surpassed the previously best-performing published method for this task, a feature fusion framework paired with a mountain gazelle optimizer that had reported 96.78 percent accuracy. The new system’s improvement to 98.43 percent represents a meaningful gain in a domain where the remaining errors often involve the most visually confusable species pairs, precisely the cases where misidentification leads to the wrong pesticide or the wrong intervention.

Accuracy alone, however, is not enough for a tool that farmers and agronomists are expected to trust. A model that achieves high scores by latching onto background clutter, such as the color of the soil or the lighting conditions under which photos were taken, will fail unpredictably in the field. To probe what the network was actually looking at, the researchers applied SHAP, or SHapley Additive exPlanations, a technique from game theory that assigns each input region a contribution score for the model’s decision. The analysis showed that the model concentrates its attention on biologically meaningful pest characteristics rather than irrelevant background regions. That transparency is a quiet but important result: it suggests the network has learned something resembling an entomologist’s eye, and it gives practitioners a way to verify that predictions rest on the insect itself.

The broader context is the economics of jute and the promise of precision agriculture. Pest infestations can significantly reduce both crop yield and fiber quality, and jute remains a cornerstone crop for Bangladesh and eastern India, where it sustains farming livelihoods and supplies the global market for natural fiber. Automated, image-based pest identification could allow farmers to detect infestations early, target interventions only where needed, and reduce the blanket pesticide spraying that carries both financial and environmental costs. The authors note that their framework achieves a practical balance between classification accuracy, computational efficiency, and interpretability, the three qualities that determine whether a laboratory model can survive deployment in an intelligent agricultural decision support system.

There are, of course, the usual caveats that separate a benchmark result from a field-ready product. The model was trained and tested on a curated public dataset, and real-world deployment will confront images taken with varied cameras, under harsh sunlight, of partially obscured insects on wind-blown plants. The researchers themselves frame the work as a foundation for practical jute pest monitoring rather than a finished deployment. Still, the trajectory is clear: a single end-to-end transformer, trained with a disciplined two-phase strategy and equipped with a fusion mechanism that preserves both texture and meaning, now classifies seventeen pest species more reliably than any method before it. For a crop that has been cultivated for millennia and identified by eye for just as long, the golden fiber may soon have a digital sentinel watching over it.

Subject of Research: Deep learning-based automated recognition of jute pest species for sustainable agriculture

Article Title: A hierarchical feature fusion transformer for robust Jute Pest recognition in sustainable agriculture

Article References: Vishwakarma, A. K., & Ranjan, R. (2026). A hierarchical feature fusion transformer for robust Jute Pest recognition in sustainable agriculture. International Journal of Machine Learning and Cybernetics, 17(10), Article 494. https://doi.org/10.1007/s13042-026-03330-y

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03330-y

Keywords: jute pests, vision transformer, MaxViT, feature fusion, deep learning, precision agriculture, transfer learning, SHAP explainability, image classification, crop protection, computer vision, sustainable farming